IDCite: A Large-Scale Multidisciplinary Citation Intent Dataset for Scholarly Knowledge Discovery

IDCite IDCite is a large-scale multidisciplinary citation dataset designed to support research on citation contexts, citation intent, scholarly information retrieval, citation recommendation, scientometrics, and scholarly knowledge graphs. Unlike conventional citation datasets that primarily represent citations as links between papers, IDCite is organized around citation events connecting citing papers to highly cited seed papers while preserving semantic and bibliographic evidence. Each citation event can incorporate citation contexts, automatically generated citation-intent annotations, publication metadata, disciplinary information, and normalized scholarly entities. The current Version 3 release contains: - 1,857,503 citation events- 1,467,045 citing papers- 23,479 highly cited seed papers- 21 Essential Science Indicators (ESI) fields- 31 observed citation-intent labels- 3,418,433 scholarly knowledge graph nodes- 6,855,117 scholarly knowledge graph edges The resource additionally provides normalized representations of authors, affiliations, journals, scientific fields, cities, countries, and citation intents. Relationship to Previous Releases IDCite represents the third stage in the evolution of the scholarly citation resources distributed through this Zenodo record. Version 1 — MDCite MDCite was released as the original large-scale multidisciplinary citation-context dataset. It focused primarily on citation contexts and citation-intent information for citation-aware scholarly analysis. Version 2 — MDContextCite and EdgeCite MDContextCite extended the initial resource into a broader scholarly citation database and dissemination framework. Within MDContextCite: - MDCite served as the core citation-context dataset.- EdgeCite provided an additional retrieval-oriented variant designed for leakage-controlled large-scale citation retrieval experiments. Version 3 — IDCite IDCite is the current structured multidisciplinary dataset release. IDCite should not be interpreted as a simple renaming of MDContextCite. Instead, it reorganizes and extends the underlying scholarly citation resources around citation events, normalized scholarly entities, and an ontology-ready scholarly knowledge graph representation. Previous MDCite and EdgeCite artifacts remain available through the Zenodo version history and provide provenance for the evolution of the resource. Data Sources IDCite integrates information from multiple scholarly infrastructures. 🔹 Scopus Scopus bibliographic records provide publication metadata, journal-level records, citation counts, and information used during highly cited seed-paper selection. 🔹 Web of Science / Journal Citation Reports Web of Science subject categories and the 2024 Journal Citation Reports are used for multidisciplinary journal-stratified sampling. One representative Web of Science category is associated with each of the 21 ESI fields, and five representative Q1 journals are selected within each category. 🔹 OpenAlex OpenAlex supports citation-link retrieval, DOI/title resolution, and structured scholarly metadata interoperability. 🔹 Semantic Scholar Semantic Scholar provides citation-linked scholarly information and reference-linked citation-context evidence associated with citing papers. 🔹 Automatic Citation Intent Classification Citation-intent annotations are generated automatically at scale using the SynIntent classifier. The resulting annotations should be interpreted as model-derived semantic annotations / weak semantic supervision, rather than manually curated gold-standard labels for every citation event. Dataset Construction Pipeline The IDCite construction workflow consists of the following major stages: 1. Journal-stratified sampling across 21 ESI fields and 21 representative Web of Science categories.2. Selection of five representative Q1 journals per category, resulting in 105 journals.3. Independent selection of the top 5% most-cited papers within each journal to construct the seed-paper corpus.4. Citation-link harvesting and citation-event construction between citing papers and seed papers.5. Citation-context collection and automatic citation-intent annotation.6. Normalization of publications, journals, authors, affiliations, geographic entities, scientific fields, and citation intents.7. Conversion of normalized entities and relationships into an ontology-ready scholarly knowledge graph. The journal-stratified sampling strategy is intended to preserve disciplinary heterogeneity and reduce the dominance of citation-intensive disciplines that would arise from applying a single global citation threshold. Because the seed corpus focuses on highly cited publications, IDCite primarily characterizes citation behavior around influential scholarly works and should not automatically be interpreted as representative of citation behavior surrounding moderately or rarely cited publications. Citation Intent Representation IDCite contains 31 observed citation-intent labels, consisting of 30 valid observed categories and one missing label retained for reproducibility. These labels include seven canonical citation intents: - background- uses- similarities- differences- motivation- extends- future work Composite labels are also retained in the released data. Composite intent labels can be decomposed into their constituent canonical intents, allowing a citation event to contribute to more than one semantic intent category. Because citation-intent labels are automatically inferred, annotation quality is model-dependent and may vary across disciplines and citation contexts. Human Evaluation of Citation-Intent Annotations For the reliability of the automatically generated citation-intent annotations, the release includes a human-evaluation archive: - `human_evaluation_Tuan_Anh_Phan.zip` This archive contains human evaluation results for citation-intent predictions on a representative multidisciplinary sample of the corpus. The evaluation is organized across 21 disciplinary categories and the 7 canonical citation intents: `background`, `uses`, `similarities`, `differences`, `motivation`, `extends`, and `future_work`. For each discipline and citation-intent category, a separate CSV file is provided. Each CSV contains evaluated citation contexts and the following three fields: - `citation context` — the citation-context text evaluated- `Prediction from classification model` — the citation intent predicted by the automatic classification model- `Prediction from human` — the corresponding citation intent assigned through human evaluation In total, the archive contains 147 CSV evaluation files (21 disciplines × 7 canonical intents), providing direct model–human comparison data across disciplinary and citation-intent categories. The human-evaluation files complement the benchmark-based evaluation of the citation-intent classifier and provide additional evidence for assessing the quality of the automatically inferred semantic annotations released with IDCite. IDCite Dataset Files The Version 3 release provides 17 Parquet files covering citation events, publications, normalized scholarly entities, and graph representations. Citation Event Files - `citation_events.parquet`- `citation_events_enriched.parquet`- `citation_events_normalized.parquet` These files contain 1,857,503 citation events at different levels of enrichment and normalization. Citing Paper Files - `citing_papers.parquet`- `citing_papers_normalized.parquet` These files contain metadata for 1,467,045 citing papers. Seed Paper Files - `seed_cited_papers.parquet`- `seed_cited_papers_normalized.parquet` These files contain metadata for 23,479 highly cited seed papers. Normalized Scholarly Entity Files - `authors.parquet`- `affiliations.parquet`- `affiliation_geo.parquet`- `journals.parquet`- `fields.parquet`- `intents.parquet`- `cities.parquet`- `countries.parquet` These lookup tables provide dataset-level normalized identifiers for scholarly entities and support consistent joins across the IDCite files. Scholarly Knowledge Graph Files - `kg_nodes.parquet`- `kg_edges.parquet` The graph representation contains: - 3,418,433 nodes- 6,855,117 typed edges Scholarly Knowledge Graph IDCite provides a heterogeneous scholarly knowledge graph derived from the normalized citation-event infrastructure. The graph represents entities including: - citation events- citing and cited publications- authors- affiliations- journals- scientific fields- cities- countries- citation intents Typed relations encode citation-event linkage, authorship, affiliation, publication venue, geographic association, disciplinary assignment, and citation-intent information. The knowledge graph is provided as a supplementary representation of IDCite and does not replace the tabular citation-event files. Ontology-ready Representation In IDCite, ontology-ready means that major scholarly entities are represented using normalized dataset-level identifiers and connected through explicit typed relationships. The term does not imply that a single external formal ontology has already been imposed on every entity. Instead, IDCite provides a structured graph-ready representation that can be mapped to or extended with formal scholarly ontologies. Intended Research Uses IDCite is designed to support research including: - citation context analysis- citation intent analysis and classification- scholarly information retrieval- citation recommendation- scientometric and bibliometric analysis- cross-disciplinary citation analysis- scholarly knowledge graph construction and analysis- graph representation learning- link prediction- research trend analysis- scholarly knowledge discovery The combination of contextual semantic evidence and structured scholarly relationships enables analyses that cannot be performed using paper-to-paper citation links alone. CitationHub CitationHub is the i